This project evaluates and compares three distinct machine learning approaches for image classification using the Caltech-256 dataset. The models tested are:
- ✅ EfficientNet-B0 (CNN)
- ✅ Support Vector Machine (SVM) with Bag of Features and SURF/SIFT descriptors
- ✅ Bernoulli Naive Bayes on simplified pixel/feature representations
Conducted as part of a group assignment for a university course on applied machine learning.
Alex Conroy, Jordan McCallum, Nebojsa Ajdarevic
Group project (QUT)
Alex’s contribution:
- Research and model validation
- Troubleshooting model performance
- Writing and editing of the final report
The goal of this project was to explore:
- How different ML approaches handle image classification
- Tradeoffs between deep learning, traditional computer vision, and naive statistical models
- Reproducibility and interpretability of each pipeline
assignment2-object-recognition/ ├── data/ # Placeholder – Caltech-256 images ├── notebooks/ │ ├── Assignment2Training.ipynb # Main training & testing logic │ └── Bernoulli-Naive-Bayes.ipynb # Additional model comparison ├── report/ │ ├── Proposal.pdf │ └── Report-v4.docx # Final group report ├── images/ │ └── image-7.png # EfficientNet architecture ├── src/ │ └── model_pipeline.py # Placeholder for reusable ML logic ├── requirements.txt # Required Python packages └── README.md
| Model | Technique | Feature Type | Accuracy | Pros | Cons |
|---|---|---|---|---|---|
| EfficientNet-B0 | CNN (Transfer Learning) | Full Image (Tensor) | High | Powerful, automated feature learning | Requires GPU, longer training |
| SVM + SURF/SIFT | Traditional CV + Bag of Features | Keypoint descriptors | Medium | Interpretability, compact | Requires tuning, not end-to-end |
| Bernoulli NB | Statistical Classifier | Binary input features | Low | Fast, simple | Low performance, crude features |
- Caltech-256:
30,000+ images across 256 categories.
Link to download
Dataset not included due to size. Please download manually and extract into ./data/.
Wrap pipelines into reusable functions/modules in src/
Explore additional model architectures (e.g. ResNet, ViT)
Test on more balanced or augmented datasets
Improve reporting and deployment readiness
This project was developed as part of an academic course. Caltech-256 dataset is publicly available for research purposes.
This project demonstrates:
- Clear ML modeling tradeoffs
- A working understanding of deep learning vs. classical CV
- Group-based scientific investigation
- Strong documentation and reproducibility